NAIL-RAPM v1.3.1: Pruned Additive Profile
NAIL-RAPM v1.3.1 is a controlled parsimony ablation of v1.3. It retains the same forward player prior, value-conditioned aging, exposure-gated cold starts, gap-returner bridge, profile shrinkage, five non-additive lineup terms, ordinary linear Ridge context fit, and regularization. The only differences are two removed additive player-profile coordinates.
Feature Change
| Additive profile feature | v1.3 | v1.3.1 |
|---|---|---|
| Three-point attempts / 100 | Included | Removed |
| Usage / 100 | Included | Removed |
| Three-point makes / 100 | Included | Included |
| Unassisted three-point makes / 100 | Included | Included |
| All other additive and non-additive terms | Included | Included |
The rationale is collinearity, not a claim that shot volume or usage are unimportant. FTA volume, three-point makes, unassisted rim makes, unassisted three-point makes, assists, and turnovers already encode much of the workload and self-creation signal. In the v1.4 partial-effect audit, usage was insufficiently resolved and three-point attempts had resolved evidence in only 7 of 29 completed posterior states. The experiment asks whether retaining those two conditional partial effects improves frozen prediction enough to justify their complexity.
Retained Coefficient Trajectories
The chart below extracts the standardized Ridge coefficient for each of the ten retained additive profile terms from every completed source-season model. Each point is a conditional effect of a one-standard-deviation home-minus-away lineup differential, holding the other profile and non-additive lineup terms fixed. It is not a standalone player rating or a causal estimate.
The completed 2022-23 source model, for example, was used to forecast the
frozen 2023-24 season; no target-season outcomes enter the corresponding
coefficient. The 10-panel chart intentionally omits only the two ablated
coordinates, three_pa_per_100 and usage_per_100.
The retained terms are directionally stable in most source seasons. Assists, three-point makes, and steals retain the same sign in all 29 source models; free-throw attempts, blocks, turnovers, and offensive-rebound claim are also consistently signed. Unassisted three-point makes remain the least stable retained term, which keeps that coordinate a candidate for a future focused ablation rather than a settled profile component.
Evaluation Contract
The candidate is fit recursively from 1996-97 through 2025-26. The same completed source-model boundary is replayed into frozen 2023-24, 2024-25, and 2025-26 regular seasons and their playoffs. v1.3 is the direct incumbent, because this is strictly a two-feature ablation rather than a comparison with an older NAIL release.
The non-promotion gate is unchanged: for the primary full-game margin RMSE, the upper endpoint of the paired 95% game-block bootstrap interval for v1.3.1 minus v1.3 must be no greater than 0.5% of v1.3's RMSE in the pooled sample and each individual frozen season.
Results
The full recursive fit, frozen replay, and paired bootstrap completed. The two-feature removal is extremely close to v1.3 and is modestly worse on most aggregate regular-season measures. It is also simpler, and its paired full-game error is within the agreed practical-harm limit in the pooled sample and every frozen season.
| Model | Regular poss. RMSE | Regular full-game RMSE | Winner accuracy | Team NetRtg RMSE | Pythagorean-win RMSE | Playoff poss. RMSE | Playoff eligible-game RMSE |
|---|---|---|---|---|---|---|---|
| NAIL-RAPM v1.3 | 1.197987 | 14.298111 | 68.44% | 3.3320 | 7.0619 | 1.192630 | 16.5034 |
| NAIL-RAPM v1.3.1 | 1.197990 | 14.307243 | 68.36% | 3.3446 | 7.1012 | 1.192646 | 16.5248 |
Direct Bootstrap Gate
The gate is evaluated only against v1.3, because it is the parent feature contract. Positive values mean that the pruned candidate has higher full-game margin RMSE.
| Scope | v1.3.1 minus v1.3 | Paired 95% interval | Gate threshold | Gate |
|---|---|---|---|---|
| Pooled | +0.0091 | [+0.0010, +0.0175] | +0.0715 | Pass |
| 2023-24 | +0.0111 | [+0.0014, +0.0207] | +0.0698 | Pass |
| 2024-25 | +0.0217 | [+0.0021, +0.0416] | +0.0727 | Pass |
| 2025-26 | -0.0044 | [-0.0160, +0.0071] | +0.0719 | Pass |
v1.3.1 becomes the preferred v1.3-branch profile contract: it removes two features whose conditional effects were redundant or insufficiently resolved, with no practically material frozen-prediction loss under the established gate. It does not replace production NAIL-RAPM v1.2, which remains stronger on the global frozen regular-season leaderboard.
Artifacts
- Recursive fit:
artifacts/models/forward_nail_rapm_v131_pruned_additive_profiles/2025-26/forward-nail-rapm-v131-pruned-additive-profiles-2025-26-20260822T053222Z-cbd60be0 - Frozen replay:
artifacts/models/nail_v131_pruned_additive_profiles_frozen_backtest/frozen_multiseason_backtest/2023-24_to_2025-26/frozen_multiseason_backtest-2023-24-to-2025-26-20260822T053952Z-3b461e0c - Bootstrap gate:
artifacts/models/nail_v131_pruned_additive_profiles_bootstrap/2023-24_to_2025-26/nail-v131-pruned-additive-profiles-bootstrap-20260822T054031Z-0391b3fe - Coefficient audit:
artifacts/models/analysis/nail_v131_pruned_additive_weight_audit/nail-v131-pruned-additive-weight-audit-20260822T123232Z-4fd27653